EasyDeepLearn

How do you prove there is no training-serving skew?

hard

Answer

  • Compare the actual feature vectors, not the code that computes them.
  • Log the features used at inference time, then take the same entities and recompute their features through the training pipeline for the same timestamp, and diff the values.
  • Any mismatch is skew, and the common causes are a transformation implemented twice in two languages, a default value that differs, timezone handling, and a training pipeline that saw data updated after the event.
  • Do this continuously on a sample rather than once, since skew appears whenever either pipeline changes.
  • The structural fix is a single feature computation path shared by both, which is the main reason feature stores exist, and shadow scoring the training pipeline against production is the strongest ongoing check.
Check yourself — multiple choice
  • Read both code paths carefully
  • Log served feature vectors and diff them against the training pipeline's values for the same entity and timestamp, continuously on a sample — and share one computation path to prevent it
  • Compare model accuracy only
  • Skew cannot be detected

Value-level diffs on the same entities are the only proof; a shared computation path is the structural fix.

#features#pipeline#data-quality

Practise MLOps & Data Quality

215 interview questions in this topic.

Related questions